Pseudo-Supervised Deep Subspace Clustering | JoVE Visualize

Area of Science:

Computer Science
Artificial Intelligence
Machine Learning

Background:

Auto-Encoder (AE)-based deep subspace clustering (DSC) methods leverage deep neural networks for powerful representation extraction.
Traditional AE-based DSC methods often overlook relational information in self-reconstruction loss, leading to suboptimal clustering performance.
Existing DSC methods face challenges in learning high-level similarities without semantic labels and incur significant memory costs due to large similarity matrices.

Purpose of the Study:

To enhance deep subspace clustering by integrating local structure information and semantic-free similarity learning.
To address the limitations of self-reconstruction loss in AE-based DSC by incorporating pairwise similarity.
To reduce the memory footprint of DSC methods and enable handling large-scale and out-of-sample data.

Main Methods:

Utilized pairwise similarity to weight reconstruction loss, capturing local structure information.
Employed a self-expression layer to learn data similarity.
Incorporated pseudo-graphs and pseudo-labels for supervised similarity learning during network training.
Implemented joint learning and iterative training strategies for an optimized solution.

Main Results:

Demonstrated the superiority of the proposed approach through extensive experiments on benchmark datasets.
Showcased improved clustering performance compared to existing AE-based DSC methods.
Validated the method's effectiveness in addressing large-scale and out-of-sample clustering problems when combined with k-nearest neighbors.

Conclusions:

The proposed method effectively captures local structure and learns high-level similarities without semantic labels, outperforming traditional AE-based DSC.
The integration of pairwise similarity, pseudo-graphs, and pseudo-labels significantly enhances clustering accuracy and representation discriminability.
The approach offers a scalable solution for large-scale and out-of-sample clustering tasks, providing a valuable advancement in the field.